Algorithms and distribution

1. What changed, precisely

For most of the twentieth century a publisher controlled both the selection and the presentation of the news it produced. A front page or a bulletin running order was an editorial statement, and every reader received the same one. Digital distribution separated those functions. Selection largely remains with the publisher; presentation and prominence increasingly belong to ranked surfaces operated by others, and now, in part, to on-site ranking operated by the publisher itself.

The consequence is subtle rather than dramatic. Editorial judgement still determines what exists, but it no longer determines what is encountered. A desk can assign its best reporter to a subject of genuine importance and reach fewer people than an offhand post, without anyone having made a decision to that effect. Understanding distribution is therefore not a marketing activity; it is part of understanding whether the journalism did what it was commissioned to do.

2. The ranking stack in outline

Nearly all production ranking systems share three layers. Candidate generation reduces an enormous catalogue to a shortlist using cheap criteria: recency, section, topic similarity, or patterns among similar readers. Scoring ranks the shortlist against a learned objective. Re-ranking then applies business and editorial constraints on top: deduplication, section caps, freshness floors, reserved slots, suppression of items under legal hold.

The distinction matters because the layers behave differently. The scoring layer is learned, opaque and hard to reason about. The re-ranking layer is written as explicit rules and can be read, argued about and changed by editors. Almost every editorial commitment a publisher can realistically make about its own ranking lives in that third layer, which is why it deserves far more attention than it usually receives.

3. Signals and the behaviour they reward

A ranked surface can only use what it observes. Observation is dominated by implicit behaviour — opens, dwell, scroll, returns — because it is abundant and requires nothing of the reader. That abundance is also the problem: implicit behaviour records what was hard to resist rather than what was valued. A headline that creates an unresolved question outperforms one that answers it, and a system learning from clicks will conclude that readers prefer the first.

Explicit signals — following, saving, dismissing, subscribing to a newsletter — are much better evidence of preference and much scarcer. Systems that weight them heavily behave more like an editor and less like a slot machine, but they need deliberate design: readers must be asked, the controls must be easy to find, and the settings must be revisited, because a topic chosen two years ago is weak evidence about a reader today.

4. Search surfaces and generated answers

Search has historically been the most valuable distribution channel for reference and explanatory journalism, because it matches an article to an expressed need rather than to a browsing mood. The arrival of generated answer panels changes that arrangement in a way that is still unfolding: a summarised response may resolve the reader’s question without a visit, while attribution and link placement vary by surface and jurisdiction.

Two responses are worth distinguishing. The first is defensive and largely technical: clear structure, honest headings, factual precision, and machine-readable markup that describes the page accurately. The second is editorial: publish material whose value is not reducible to a summary — original reporting, documents, interviews, data the reader can interrogate, and analysis that depends on judgement. The second is slower and considerably more durable.

5. Measuring concentration

Reach figures conceal how unevenly attention is distributed. Almost every publication finds that a small number of items absorb most of the reading, and the size of that share is a meaningful editorial fact. If the top five items take seventy per cent of impressions, the effective front page is five items long, whatever the site’s design suggests.

Track all three over time rather than at a single moment. A rising top-five share with a falling catalogue reach describes a system that is progressively narrowing what the audience encounters, which is a decision the newsroom should make knowingly rather than discover a year later.

6. Owned channels and what they are for

Owned channels are not merely a hedge against platform change; they behave differently as editorial products. A newsletter has a fixed sequence and a single editor’s judgement, so it can carry an item that would never survive ranking. A podcast feed reaches subscribers regardless of an algorithm’s opinion. A well-maintained site feed lets other systems, including automated ones, discover the full output rather than a ranked slice.

The trade is scale. Owned channels grow slowly and require sustained editorial attention, and a newsletter published irregularly is worse than none. The realistic target is not to replace ranked distribution but to ensure that some meaningful proportion of the audience can be reached without permission from anyone, and to know what that proportion is.

7. Diversity of exposure as an editorial commitment

Publishers routinely say they serve a broad public. Ranked distribution makes that claim testable. Measure the share of readers who encountered at least one public-interest item in a week; the number of distinct sections a typical reader was exposed to; the proportion of local coverage that reached anyone outside its own locality. These are ordinary counts, and they can be reported internally each month without any additional data collection about individuals.

Where the numbers are poor, the fix is usually structural rather than exhortatory: a reserved slot, a floor, an exploration budget for new items, a cap on how much of any surface one section can occupy. Rules survive busy weeks; intentions do not.

8. Transparency about your own ranking

The Digital Services Act requires providers of certain online platforms to explain, in plain and intelligible language, the main parameters used in their recommender systems and any options for modifying them, with additional duties for very large platforms and search engines. Most publishers are not in scope for those specific obligations, but the underlying practice is cheap and useful.

A short public page can state what the site’s ranking considers, what it deliberately ignores, which positions are fixed by editors, and how a reader can obtain a chronological view. Publishing it has an internal benefit as well: writing the description forces someone to establish what the system actually does, which is frequently not what the documentation claims. Confirm your own legal position with a qualified professional rather than inferring it from a general description.

9. Keeping a distribution record

Distribution decisions are usually made quickly and remembered badly. A simple log solves both problems: one row per significant change — an objective adjustment, a new reserved slot, a platform policy change, a format experiment — with the date, the reason and the expected effect. Reviewed quarterly against the concentration and diversity figures, it converts a series of tactical moves into evidence about what actually works.

The log also matters for accountability. When a reporter asks why their coverage reached fewer people than expected, a dated record of what changed is a far better answer than a shrug, and it occasionally reveals that the system, not the journalism, was the variable.

10. A quarterly distribution review

Run this as a fixed exercise rather than a response to a bad month. Each step produces a short written output, and the whole cycle should take a few hours spread over a fortnight.

1. Step 1 — restate the objective Write in one sentence what the on-site ranking currently maximises, and check that it matches the configuration.
2. Step 2 — measure concentration Calculate top-five share, catalogue reach and exposure overlap for the period, and compare against the previous quarter.
3. Step 3 — test the floors Confirm that reserved editorial slots and section caps were applied on a sample of days, not merely configured.
4. Step 4 — audit external dependency Record the share of sessions from each external source and the change since the last review.
5. Step 5 — review the log Read every distribution change made in the quarter and note which had the expected effect.
6. Step 6 — update the public statement Amend the ranking description so it still matches practice, and date the revision.

Distribution review checklist

• The on-site ranking objective is written in one plain sentence and dated.
• Re-ranking rules, reserved slots and caps are documented and readable by editors.
• Top-five share, catalogue reach and exposure overlap are measured each quarter.
• New items receive an exploration budget rather than competing on history alone.
• A holdout or chronological view exists as an internal comparison baseline.
• The share of sessions from each external source is tracked monthly.
• At least one owned channel can reach a known proportion of the audience directly.
• A public page describes what the ranking uses and how to obtain a chronological view.
• Every distribution change is logged with a date, a reason and an expected effect.
• Legal questions about platform obligations go to a qualified professional.

The purpose is not to optimise any single figure. It is to ensure that nobody can say, a year from now, that the newsroom did not know what its own distribution was doing.

Questions readers ask